Security

Understanding Akira's security model and isolation guarantees.

Threat Model

Akira is designed for one primary threat: running untrusted code safely.

This includes:

• AI-generated code that might be malicious or buggy
• User-submitted code in multi-tenant environments
• Third-party scripts and tools
• Experimental code during development

Isolation Layers

Akira provides multiple layers of isolation:

┌─────────────────────────────────────────────────────┐
│                   Your Infrastructure                │
├─────────────────────────────────────────────────────┤
│                  Akira Platform                      │
│  ┌───────────────────────────────────────────────┐  │
│  │            Hardware Virtualization             │  │
│  │  ┌─────────┐  ┌─────────┐  ┌─────────┐       │  │
│  │  │ Sandbox │  │ Sandbox │  │ Sandbox │       │  │
│  │  │  (VM)   │  │  (VM)   │  │  (VM)   │       │  │
│  │  │         │  │         │  │         │       │  │
│  │  │ Process │  │ Process │  │ Process │       │  │
│  │  │ + Code  │  │ + Code  │  │ + Code  │       │  │
│  │  └─────────┘  └─────────┘  └─────────┘       │  │
│  └───────────────────────────────────────────────┘  │
└─────────────────────────────────────────────────────┘

Layer 1: Hardware Virtualization

Each sandbox runs in its own micro-VM with hardware-level isolation:

Layer 2: Minimal Attack Surface

Micro-VMs have a tiny attack surface compared to traditional VMs:

• Minimal device emulation
• No legacy hardware support
• Reduced hypervisor interface
• Purpose-built for sandboxing

Layer 3: Resource Limits

Bounded resources prevent abuse:

• CPU limits - Maximum cores allocated
• Memory limits - Maximum RAM available
• Time limits - Automatic termination after timeout
• Storage limits - Bounded disk space

What Sandboxes CAN'T Do

A sandbox cannot:

What Sandboxes CAN Do

By default, sandboxes can:

• Execute arbitrary code
• Make outbound network requests
• Read/write their own filesystem
• Use allocated CPU and memory
• Access sandbox storage

Network Security

Default Configuration

• Outbound: Allowed (can reach internet)
• Inbound: Blocked (cannot be reached)
• Inter-sandbox: Blocked

Network Isolation

Sandboxes cannot:

• Communicate with other sandboxes
• Access internal network services
• Scan or probe your infrastructure

API Security

Authentication

All API requests require authentication via x-api-key header:

curl -H "x-api-key: YOUR_API_KEY" https://api.akiralabs.ai/v1/sandboxes/list

Key Management

• Keys are scoped to your organization
• Rotate keys regularly
• Never commit keys to version control
• Use environment variables

// Good: Environment variable
const client = new SandboxSDK({
  apiKey: process.env['AKIRA_API_KEY'],
});
 
// Bad: Hardcoded key
const client = new SandboxSDK({
  apiKey: 'sk_live_abc123', // Never do this!
});

Best Practices

For AI Agent Developers

1. Always use sandboxes - Never run AI-generated code directly
2. Set timeouts - Prevent infinite loops and runaway processes
3. Validate outputs - Don't trust sandbox outputs blindly
4. Clean up - Delete sandboxes when done

For Platform Operators

1. Rotate API keys - Regular key rotation
2. Monitor usage - Watch for anomalous patterns
3. Set resource limits - Appropriate CPU/memory bounds
4. Review audit logs - Track sandbox operations

For Security-Sensitive Workloads

1. Disable network - Block outbound if not needed
2. Use ephemeral sandboxes - Don't persist state
3. Limit storage access - Minimize data exposure
4. Implement timeouts - Short execution limits

Comparison: Containers vs Akira

Security Guarantees

Akira provides:

• Confidentiality - Sandboxes cannot read other sandboxes
• Integrity - Sandboxes cannot modify host or other sandboxes
• Availability - Resource limits prevent DoS

Akira does NOT protect against:

• Malicious use of your own API keys
• Data exfiltration via network (if enabled)
• Side-channel attacks (theoretical)

Incident Response

If you suspect a security issue:

1. Revoke affected API keys immediately
2. Delete suspicious sandboxes
3. Review audit logs
4. Contact security@akiralabs.ai

Next Steps

• Learn about sandbox configuration
• Explore use cases for coding agents
• Review the API reference